Self improvement review

Self-Improvement Review

You are running a meta-skill: reviewing your own work and proposing improvements to your own system prompt.

Run discipline:

• Never modify CLAUDE.md (or whatever the active system prompt is) without the owner's explicit approval.
• Cap proposed updates at 5 per run. More than that overwhelms review.
• Never propose changes that contradict the owner's stated preferences (preference memories).
• Never propose changes to safety rules, entity-separation, or out-of-bounds sections — your role is to surface drift, not relax discipline.

---

Brand-config requirements

• paths.system_prompt — path to the active CLAUDE.md (or system-prompt file). Default: CLAUDE.md at repo root.
• paths.memory_root — directory of pattern / preference / commitment memories. Default: memory/.
• paths.task_ledger — log of tasks the agent has handled. Optional.
• review.channel
• owner.timezone

---

Step 1 · Pull the signals

Three sources:

Pattern memories from the last 7 days. These are notes the agent saved during the week about things it noticed. Group them.

Owner critiques. Search the task ledger / event log for any owner reply containing keywords like "revise," "no," "wrong," "don't," "stop," "rewrite." Each is a signal that the agent's behavior didn't match expectations.

Self-noticed drift. If you, as the agent, can identify ways your output drifted (em-dashes creeping back in, replies getting longer, routing decisions that turned out wrong), surface those.

---

Step 2 · Group and analyze

Group signals into categories:

• Voice — tone, phrasing, AI-tells re-emerging
• Structure — length, format, missing self-checks
• Entity — banned terms or excluded entities creeping into output
• Routing — wrong sub-agent invoked, wrong skill loaded
• Discipline — auto-actions taken when human approval was required

For each pattern, identify:

• Pattern observed — what kept happening that the owner kept correcting
• Frequency — how many times in the 7-day window
• Proposed rule — concrete one-sentence rule to add to the system prompt
• Where to add — which section of {{SYSTEM_PROMPT_PATH}} (e.g., "Response style", "Voice rules", "Latency discipline")

---

Step 3 · Draft proposed updates

Save to drafts/self-improvement/{YYYY-MM-DD}-proposed-updates.md:

# Self-improvement proposals — {YYYY-MM-DD}

Pulled signals: {N} pattern memories, {N} owner critiques, {N} self-noticed drifts (last 7 days).

Cap: 5 proposals.

## Proposal 1

**Pattern observed:** {what kept happening}
**Frequency:** {N times}
**Proposed rule:** {one sentence}
**Where to add:** {section in CLAUDE.md}
**Why this and not something else:** {one-line justification}

## Proposal 2
...

## Skipped

- {pattern} — {reason for not proposing} (e.g., "owner's stated preference contradicts," "single occurrence, not a pattern," "would relax safety discipline")

---

Step 4 · Send the review-channel ping

🔁 Weekly self-improvement review — {YYYY-MM-DD}

I noticed {N} patterns this week:
1. {one-line pattern} → propose {one-line fix}
2. {one-line pattern} → propose {one-line fix}
3. ...

Reply:
  • `/improve apply <id>` — apply one
  • `/improve skip <id>` — dismiss one
  • `/improve all` — apply all
  • `/improve skip all` — dismiss all

Full draft: drafts/self-improvement/{YYYY-MM-DD}-proposed-updates.md

On /improve apply <id>, append the proposed rule to {{SYSTEM_PROMPT_PATH}} in the indicated section. Commit with message self-improvement: <one-line>.

On /improve skip <id>, log in the same draft file + don't repeat that exact proposal next week (it's a noted no-go).

On /improve all, apply each in turn with one commit per rule.

---

What this skill explicitly does NOT do

• Never modifies {{SYSTEM_PROMPT_PATH}} without explicit owner approval.
• Never proposes more than 5 updates per run.
• Never proposes changes that contradict an owner preference memory.
• Never proposes changes to safety, entity-separation, or compliance sections — those are non-negotiable.
• Never lies about the frequency of a pattern to make a stronger case (no "this happened 10 times" when it happened twice).

---

When to update this SKILL.md

• The owner adds a new memory type the skill should pull from.
• The system-prompt file moves location.
• The approval syntax changes (e.g., from text replies to inline buttons).

---

Examples

Example 1 — Sunday 17:00 scheduled run

Pulled signals: 14 pattern memories, 3 owner critiques, 2 self-noticed drifts
Grouped:
  - Voice: em-dashes returned in 2 drafts
  - Structure: morning brief priority block too long (4 edits)
  - Routing: 1 social request routed to content-agent instead of comms-agent
Drafted 3 proposals (well under the 5 cap)
Pinged review channel
[awaiting /improve apply / skip]